Unverified50% confidenceOpinionExact time
智能体的注意力资源围绕奖励信号分配，只有与奖励相关的信息才会在反向传播中获得有效梯度并被网络编码
1
Sources
50%
Confidence
Long-term
Relevance
8/9/2026
First Seen
Sources
Related Claims
Unverified人类标注者在评估答案时存在信息越丰富越好的认知偏差，导致奖励模型将篇幅与质量隐性挂钩，是模型冗余的成因之一72% similarUnverified元认知反馈通过奖励置信度与准确率的一致性，在奖励信号中引入认知诚实度这一与内容质量正交的新维度67% similarUnverified注意力机制本质上是一种动态加权求和，Query代表在找什么、Key代表能提供什么、Value代表实际内容67% similarUnverifiedBilibili's recommendation algorithm uses user behavior signals such as likes, coins, favorites, and comments as the core weights for content distribution, with favorites and coins weighted higher than likes.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/718396API
curl https://kongchang.com/api/v1/knowledge/claims/718396MCP
get_claim(id=718396)